Overview of PI-SDE. (a) PI-SDE combines neural SDEs and physics-informed insights to learn cellular dynamics from time-series scRNA-seq data. Specifically, we depict cellular differentiation as diffusion processes governed by SDEs with both drift and diffusion coefficients modeled through separate neural networks. By integrating HJ regularization, the resulting physics-informed loss function guides our model to find the optimal potential landscape with accurate prediction and biological interpretability. PI-SDE takes the input of time-series scRNA-seq data and outputs (b) the predicted unseen data, (c) the reconstructed energy potential landscape that resembles the original Waddington landscape, and (d) the inferred cellular velocity.